As I write, the world is gripped in shockwaves of hurt and anger about the recent death of George Floyd and the issue of racism.
And many are asking what effective altruism (or EA) has to say about this (see here, or here, or here)
I have engaged with the EA community for many years, but I don’t consider myself an authoritative voice on effective altruism.
I am a non-white person living in a predominantly white country, but I don’t consider myself an authoritative voice on racism.
But I wanted to share some thoughts.
First of all, let’s just acknowledge that the EA community has been—in its own way—fighting against a particular form of racism right from the earliest days of the EA movement.
What greater racism is there than the horrifically uneven distribution of resources between people all because of an accident of their birth? And how disgusting that some of the worst off should be condemned to death as a result?
Is not the obscene wealth of the developed world in the face of tractable, cost-effective ways of saving lives in the developing world wholly unjustifiable if we were treating people as equals, regardless of where they are, and regardless of the colour of their skin?
I still find this argument compelling. And I would encourage the EA movement to be proud of what it has done, proud of the hundreds of millions of dollars already moved in an expression of global solidarity to people around the world.
But in some ways this argument feels insular.
Is EA really all about taking every question and twisting it back to malaria nets and AI risk?
Those of us who, like me, have spent most of their lives in the UK and are old enough will remember the name of Stephen Lawrence. And for those of us, like me, who have spent a substantial chunk of their lives in charities in South East London, his name will have followed us like a ghost.
For those who have sensed or lived systemic racism, I do think that the EA way of thinking has something to offer. And something more than “donate to the Against Malaria Foundation”.
In this post, I set out some thoughts. I would love to have provided good solutions, like “this is the best place to donate” or “this is the best thing to do” but the range of existing thoughts on this topic is too broad and complex for me to be able to do that now.
I think the most important thing that an EA mindset has to offer is this:
EA is not just about finding the right answers, it’s about asking fundamental questions too.
The effective altruism movement is unusual. Not only do EA-minded people answer questions like “what is the best way to improve global health” (finding the right answers). The EA approach also poses questions like “what is the best cause area to tackle, is it global health, or is it existential risk, or is it something else?” (asking more fundamental questions).
At first glance, asking the more fundamental questions about cause prioritisation risks subverting our goal. We may conclude that tackling an intractable thing like systemic racism isn’t really the best bang for your buck, and then we’re back to turning everything into malaria nets again.
On second glance, it’s clear that EA does have something to offer.
For example, someone who cares about animals would be encouraged by an effective altruist to consider the different “sub-causes” within the animal cause area, and provided data and arguments about why some are much more effective than others.
So in that vein, here are some thoughts about tackling systemic racism as seen through (my interpretation of) an EA lens.
These thoughts will raise more questions than answers. My hope and intention is that these are good, useful questions.
Before I get started, just an observation: achieving change is hard. Especially in a tough area like this. It will require substantial amounts of financial input and lots of capable dedicated people putting in huge amounts of effort, as well as good collaboration between them. What follows is thinking to help prioritise and make decisions about that effort.
Most of the rest of this post may feel a bit technical. Readers who don’t want to get caught up in my proposed approach are encouraged to skip to the section headed “Final thoughts”.
--------------------------
The first step is to work out what we want to achieve. My understanding of what we want is to tackle systemic racism at a national level (e.g. in the US, or the UK).
Note that this goal does not presume that the only outcomes we care about are at the level of the police. It’s possible for George Floyd’s death to be the clarion call, but for the ripples to reach far and wide beyond the judicial system.
I start by distinguishing between
direct social interventions
population-wide campaigning to change social norms, and
lobbying work to change policy.
1. Direct social interventions
In finding direct social interventions, I distinguish between
We have the evidence, but nobody is implementing it
People are doing things, but the evidence for their effectiveness is weak or even non-existent
(Projects that people are working on *and* which have strong evidence are much rarer than we might hope)
In both categories, I suggest we consider:
What does the evidence say?
How cost-effective do we think it’s going to be?
This table illustrates the sort of thinking I would propose. For the avoidance of doubt, this is not a real assessment, it’s just an illustration to show what a real assessment might look like.
Quality of evidence
The effective altruism movement tends to have a thread of scepticism about the effectiveness of most social interventions (see, e.g. this post by my organisation SoGive or this post by 80,000 hours).
If you are convinced by this, then you want social interventions which have a good evidence base behind them.
I did a quick search for RCTs on tackling racism. The only relevant piece of evidence I could find was Platz et al 2017, which found that diversity training for police recruits was effective at stopping a decline in values. However this was based on a not-particularly-large sample size and I couldn’t find any more studies in the google scholar search, the papers cited by Platz et al 2017, or in the papers which cited Platz et al 2017.
More evidence may exist (my search was very quick).
Once we have the evidence, we assess it for its quality. This includes considering the quality of controls used (randomised is generally better), sample size, risk of file drawer effects, risk of moving the goalposts, risks of HARKing (hypothesising after the results are known), risks from studies not being blinded (where applicable).
If the intervention is being done by an existing charity, the work is often fairly weakly evidenced (it’s more common for someone to start a charity first and then seek evidence than the other way round, and the incentives for robust evidence are not strong once a charity is already running).
The effective altruist tendency to focus on evidence stems from a desire to genuinely achieve change.
Cost-effectiveness
The effective altruism movement loves cost-effectiveness. This isn’t stinginess; it’s because if the funds available are finite (which they always are) then you want to do as much good as you can with the money available.
Estimating cost-effectiveness of an intervention is often hard. Especially when there is not an established organisation already doing the intervention.
An organisation I founded called SoGive has developed techniques for estimating the cost per output of an existing charity intervention.
And for situations where this data doesn’t exist, here is a sketch of an approach for a very quick-and-dirty cost-effectiveness model:
Assume that the cost per intervention is £500 (source: a review of the SoGive database of cost per output across several charitable interventions/activities working directly with people in the UK suggests a cost between £100 and £3,000 is likely; £500 is a representative figure (technically, it’s the “geometric mean”))
Build a model of the outcomes suggested by the evidence (e.g. Notional funding amount of £10k → 20 HR departments supported with unconscious bias training → 100 cases of people getting a job they otherwise wouldn’t have got) This could be turned into a spreadsheet model with confidence intervals including pessimistic and optimistic estimates
These estimates of the cost per outcome can be compared with each other
This analysis can help to narrow down a shortlist of more effective interventions for people to work on / fund.
2. Campaigning to change social norms
For changing social norms and behaviours, the most well-evidenced intervention that I know of is the Saturation + model followed by Development Media International (DMI). This is essentially about exposing the target audience to a high frequency of messaging. This is not dissimilar to applying the power of advertising to social change.
To my knowledge, evidence for the Saturation + model exists when it comes to improving public health, but not for changing social norms like racism. It would certainly need a new evidence base before its effectiveness in this context were established, however the existing evidence should at least give us hope.
I’ve only referred to Saturation + in this section. It may be that another, better model exists; a full review of other alternatives should ideally be performed.
Again, a cost-effectiveness model would be needed here. This could be determined by looking at the existing models for DMI’s work and making adjustments (and of course, it’s very hard to know how to adjust for the fact that the method is being used in a very different context to public health)
3. Lobbying to change policy
Having defined the scope as tackling systemic racism in any context, we have to consider which systems or institutions are candidates for improvement.
Before we do that, let’s remember the horrific video of George Floyd’s neck being knelt on as he lay defenceless on the floor, with voices questioning whether the police really needed to be so heavy-handed. It was brutal.
But how do we choose between tackling police brutality and the many other ways that society brutalises people of colour?
I wouldn’t claim that the below list is comprehensive, and if someone properly conducted this review, they would need to consider which rows to include.
This table illustrates the sort of thinking I would propose. For the avoidance of doubt, this is not a real assessment, it’s just an illustration to show what a real assessment might look like.
Prevalence
Prevalence considers how many people are affected by this issue. E.g. education-related racism might be prevalent (because everyone at least starts going to school) or not so prevalent (because only some schools have a racism problem). Health might have low prevalence (because some people have relatively few health needs).
This is best assessed by literally trying to size the problem; e.g. how many people in the US / UK/ wherever are suffering from this. This can be estimated from existing research on the topic, or where such research doesn’t exist or is inadequate, then new primary research may be needed.
Severity
This is a tough assessment. It’s asking how bad is it (e.g.) to be discriminated against in an educational context compared to an employment context. It could be assessed by a combination of qualitative research (e.g. reviewing interviews with people who have suffered from the problem) and quantitative research (e.g. how often do racist incidents of various levels of severity occur?)
Methods which are not really good enough, but which would be sufficient in the absence of better research , include a quick survey asking a sample of respondents to compare different outcomes for how bad they are, or for a really quick and dirty first attempt, the researcher could simply use their own judgement on what they think would be worse. While this is clearly inadequate, it is certainly better than the common alternative of making this judgement implicitly (which is inevitably what is happening when someone—consciously or otherwise—chooses to support one of these areas and not another).
Evidence that the problem is real
For any of these areas, there will be some people who believe that the racism problem either doesn’t exist, or at least is overblown.
This factor is important for two reasons:
Reason 1: If the problem isn’t real, the intervention won’t be effective
Reason 2: If the evidence is weak, it may make lobbying harder
The importance of reason 1 depends on your expectations of the skews/biases in the evidence base. If you think that publication bias, for example, leads to a skew in the evidence towards sensationalist results which indicate racism, then reason 1 is important. However it may be that the evidence base around the existence/prevalence of racism is just as likely to include the bias of wanting to show that racism is overblown/non-existent.
The techniques used above to assess the quality of evidence would largely be applicable here.
Tractability
Factors that may influence tractability include whether the institution can be changed without undergoing substantial costs, whether there are entrenched cultural norms which may invalidate any policy change, and whether there is public opinion onside.
And lastly lobbying needs a cost-effectiveness model too. This would involve looking at past lobbying efforts and estimating how much had been spent on the lobbying effort in order to achieve the policy change. This can be a fraught calculation for various reasons.
Bringing it all together
The last step is to take the cost-effectiveness estimates from the best option(s) in each of direct social interventions, campaigning to change social norms, and lobbying for policy change. Then we compare them.
When doing so, it would be important to consider the uncertainty in each area: these cost-effectiveness estimates are performed quite crudely, and should not be taken literally.
If it can be concluded from this that one area clearly outperforms another, then we have successfully narrowed our focus area.
--------------------------
Final thoughts
I wrote this partly because I think racism matters.
And I wrote it because I wanted to illustrate that EA can be a tool that can be applied in so many contexts.
It is intended to illustrate the following properties of the effective altruism approach
Being led by the heart, and not letting our heart be limited in where it takes us
George Floyd’s death might have motivated us, but that doesn’t mean that the justice system is the only place where racism is prevalent
Reason and evidence, as an expression of caring about making real change
Evidence matters because we are not devoting our time or money to these issues for tokenistic reasons, we do it because we want to effect real change
If, for the same money/time, we can help more people suffer less racism, this is great. Because every instance of racism averted counts.
The systematic approach I’ve highlighted here might seem almost callous. But to my mind it’s a sign of really caring.
Another feature of effective altruism is humility. So in that vein let me reiterate that there may be many things that I’ve got wrong here, both in terms of my thinking of what is the best thing to do for tackling racism, and in terms of capturing what EA really is.
If I were to pick one thing that’s most likely to be a mistake, it’s the omission of further research from the list of (1) direct social interventions (2) campaigning to change social norms (3) lobbying for policy change. I did this because EA isn’t just about paralysis by analysis. It’s about doing as much as it’s about thinking.
But the approach set out above is based heavily on the evidence, and if systemic racism or unconscious bias extends as far as the institutions performing the research, this jeopardises our evidence base.
And lastly our doubts about the malign influence of institutional prejudice or unconscious bias do not just reach as far as our data.
They should reach ourselves as well.
In the UK there’s a campaign group called Charity So White which tackles institutional racism within the charity sector. To finally raise a question too broad for me to answer here: Do we also need an “EA So White” too?
EA and tackling racism
As I write, the world is gripped in shockwaves of hurt and anger about the recent death of George Floyd and the issue of racism.
And many are asking what effective altruism (or EA) has to say about this (see here, or here, or here)
I have engaged with the EA community for many years, but I don’t consider myself an authoritative voice on effective altruism.
I am a non-white person living in a predominantly white country, but I don’t consider myself an authoritative voice on racism.
But I wanted to share some thoughts.
First of all, let’s just acknowledge that the EA community has been—in its own way—fighting against a particular form of racism right from the earliest days of the EA movement.
What greater racism is there than the horrifically uneven distribution of resources between people all because of an accident of their birth? And how disgusting that some of the worst off should be condemned to death as a result?
Is not the obscene wealth of the developed world in the face of tractable, cost-effective ways of saving lives in the developing world wholly unjustifiable if we were treating people as equals, regardless of where they are, and regardless of the colour of their skin?
I still find this argument compelling. And I would encourage the EA movement to be proud of what it has done, proud of the hundreds of millions of dollars already moved in an expression of global solidarity to people around the world.
But in some ways this argument feels insular.
Is EA really all about taking every question and twisting it back to malaria nets and AI risk?
Those of us who, like me, have spent most of their lives in the UK and are old enough will remember the name of Stephen Lawrence. And for those of us, like me, who have spent a substantial chunk of their lives in charities in South East London, his name will have followed us like a ghost.
For those who have sensed or lived systemic racism, I do think that the EA way of thinking has something to offer. And something more than “donate to the Against Malaria Foundation”.
In this post, I set out some thoughts. I would love to have provided good solutions, like “this is the best place to donate” or “this is the best thing to do” but the range of existing thoughts on this topic is too broad and complex for me to be able to do that now.
I think the most important thing that an EA mindset has to offer is this:
EA is not just about finding the right answers, it’s about asking fundamental questions too.
The effective altruism movement is unusual. Not only do EA-minded people answer questions like “what is the best way to improve global health” (finding the right answers). The EA approach also poses questions like “what is the best cause area to tackle, is it global health, or is it existential risk, or is it something else?” (asking more fundamental questions).
At first glance, asking the more fundamental questions about cause prioritisation risks subverting our goal. We may conclude that tackling an intractable thing like systemic racism isn’t really the best bang for your buck, and then we’re back to turning everything into malaria nets again.
On second glance, it’s clear that EA does have something to offer.
For example, someone who cares about animals would be encouraged by an effective altruist to consider the different “sub-causes” within the animal cause area, and provided data and arguments about why some are much more effective than others.
So in that vein, here are some thoughts about tackling systemic racism as seen through (my interpretation of) an EA lens.
These thoughts will raise more questions than answers. My hope and intention is that these are good, useful questions.
Before I get started, just an observation: achieving change is hard. Especially in a tough area like this. It will require substantial amounts of financial input and lots of capable dedicated people putting in huge amounts of effort, as well as good collaboration between them. What follows is thinking to help prioritise and make decisions about that effort.
Most of the rest of this post may feel a bit technical. Readers who don’t want to get caught up in my proposed approach are encouraged to skip to the section headed “Final thoughts”.
--------------------------
The first step is to work out what we want to achieve. My understanding of what we want is to tackle systemic racism at a national level (e.g. in the US, or the UK).
Note that this goal does not presume that the only outcomes we care about are at the level of the police. It’s possible for George Floyd’s death to be the clarion call, but for the ripples to reach far and wide beyond the judicial system.
I start by distinguishing between
direct social interventions
population-wide campaigning to change social norms, and
lobbying work to change policy.
1. Direct social interventions
In finding direct social interventions, I distinguish between
We have the evidence, but nobody is implementing it
People are doing things, but the evidence for their effectiveness is weak or even non-existent
(Projects that people are working on *and* which have strong evidence are much rarer than we might hope)
In both categories, I suggest we consider:
What does the evidence say?
How cost-effective do we think it’s going to be?
This table illustrates the sort of thinking I would propose. For the avoidance of doubt, this is not a real assessment, it’s just an illustration to show what a real assessment might look like.
Quality of evidence
The effective altruism movement tends to have a thread of scepticism about the effectiveness of most social interventions (see, e.g. this post by my organisation SoGive or this post by 80,000 hours).
If you are convinced by this, then you want social interventions which have a good evidence base behind them.
I did a quick search for RCTs on tackling racism. The only relevant piece of evidence I could find was Platz et al 2017, which found that diversity training for police recruits was effective at stopping a decline in values. However this was based on a not-particularly-large sample size and I couldn’t find any more studies in the google scholar search, the papers cited by Platz et al 2017, or in the papers which cited Platz et al 2017.
More evidence may exist (my search was very quick).
Once we have the evidence, we assess it for its quality. This includes considering the quality of controls used (randomised is generally better), sample size, risk of file drawer effects, risk of moving the goalposts, risks of HARKing (hypothesising after the results are known), risks from studies not being blinded (where applicable).
If the intervention is being done by an existing charity, the work is often fairly weakly evidenced (it’s more common for someone to start a charity first and then seek evidence than the other way round, and the incentives for robust evidence are not strong once a charity is already running).
The effective altruist tendency to focus on evidence stems from a desire to genuinely achieve change.
Cost-effectiveness
The effective altruism movement loves cost-effectiveness. This isn’t stinginess; it’s because if the funds available are finite (which they always are) then you want to do as much good as you can with the money available.
Estimating cost-effectiveness of an intervention is often hard. Especially when there is not an established organisation already doing the intervention.
An organisation I founded called SoGive has developed techniques for estimating the cost per output of an existing charity intervention.
And for situations where this data doesn’t exist, here is a sketch of an approach for a very quick-and-dirty cost-effectiveness model:
Assume that the cost per intervention is £500 (source: a review of the SoGive database of cost per output across several charitable interventions/activities working directly with people in the UK suggests a cost between £100 and £3,000 is likely; £500 is a representative figure (technically, it’s the “geometric mean”))
Build a model of the outcomes suggested by the evidence (e.g. Notional funding amount of £10k → 20 HR departments supported with unconscious bias training → 100 cases of people getting a job they otherwise wouldn’t have got) This could be turned into a spreadsheet model with confidence intervals including pessimistic and optimistic estimates
These estimates of the cost per outcome can be compared with each other
This analysis can help to narrow down a shortlist of more effective interventions for people to work on / fund.
2. Campaigning to change social norms
For changing social norms and behaviours, the most well-evidenced intervention that I know of is the Saturation + model followed by Development Media International (DMI). This is essentially about exposing the target audience to a high frequency of messaging. This is not dissimilar to applying the power of advertising to social change.
To my knowledge, evidence for the Saturation + model exists when it comes to improving public health, but not for changing social norms like racism. It would certainly need a new evidence base before its effectiveness in this context were established, however the existing evidence should at least give us hope.
I’ve only referred to Saturation + in this section. It may be that another, better model exists; a full review of other alternatives should ideally be performed.
Again, a cost-effectiveness model would be needed here. This could be determined by looking at the existing models for DMI’s work and making adjustments (and of course, it’s very hard to know how to adjust for the fact that the method is being used in a very different context to public health)
3. Lobbying to change policy
Having defined the scope as tackling systemic racism in any context, we have to consider which systems or institutions are candidates for improvement.
Before we do that, let’s remember the horrific video of George Floyd’s neck being knelt on as he lay defenceless on the floor, with voices questioning whether the police really needed to be so heavy-handed. It was brutal.
But how do we choose between tackling police brutality and the many other ways that society brutalises people of colour?
I wouldn’t claim that the below list is comprehensive, and if someone properly conducted this review, they would need to consider which rows to include.
This table illustrates the sort of thinking I would propose. For the avoidance of doubt, this is not a real assessment, it’s just an illustration to show what a real assessment might look like.
Prevalence
Prevalence considers how many people are affected by this issue. E.g. education-related racism might be prevalent (because everyone at least starts going to school) or not so prevalent (because only some schools have a racism problem). Health might have low prevalence (because some people have relatively few health needs).
This is best assessed by literally trying to size the problem; e.g. how many people in the US / UK/ wherever are suffering from this. This can be estimated from existing research on the topic, or where such research doesn’t exist or is inadequate, then new primary research may be needed.
Severity
This is a tough assessment. It’s asking how bad is it (e.g.) to be discriminated against in an educational context compared to an employment context. It could be assessed by a combination of qualitative research (e.g. reviewing interviews with people who have suffered from the problem) and quantitative research (e.g. how often do racist incidents of various levels of severity occur?)
Methods which are not really good enough, but which would be sufficient in the absence of better research , include a quick survey asking a sample of respondents to compare different outcomes for how bad they are, or for a really quick and dirty first attempt, the researcher could simply use their own judgement on what they think would be worse. While this is clearly inadequate, it is certainly better than the common alternative of making this judgement implicitly (which is inevitably what is happening when someone—consciously or otherwise—chooses to support one of these areas and not another).
Evidence that the problem is real
For any of these areas, there will be some people who believe that the racism problem either doesn’t exist, or at least is overblown.
This factor is important for two reasons:
Reason 1: If the problem isn’t real, the intervention won’t be effective
Reason 2: If the evidence is weak, it may make lobbying harder
The importance of reason 1 depends on your expectations of the skews/biases in the evidence base. If you think that publication bias, for example, leads to a skew in the evidence towards sensationalist results which indicate racism, then reason 1 is important. However it may be that the evidence base around the existence/prevalence of racism is just as likely to include the bias of wanting to show that racism is overblown/non-existent.
The techniques used above to assess the quality of evidence would largely be applicable here.
Tractability
Factors that may influence tractability include whether the institution can be changed without undergoing substantial costs, whether there are entrenched cultural norms which may invalidate any policy change, and whether there is public opinion onside.
And lastly lobbying needs a cost-effectiveness model too. This would involve looking at past lobbying efforts and estimating how much had been spent on the lobbying effort in order to achieve the policy change. This can be a fraught calculation for various reasons.
Bringing it all together
The last step is to take the cost-effectiveness estimates from the best option(s) in each of direct social interventions, campaigning to change social norms, and lobbying for policy change. Then we compare them.
When doing so, it would be important to consider the uncertainty in each area: these cost-effectiveness estimates are performed quite crudely, and should not be taken literally.
If it can be concluded from this that one area clearly outperforms another, then we have successfully narrowed our focus area.
--------------------------
Final thoughts
I wrote this partly because I think racism matters.
And I wrote it because I wanted to illustrate that EA can be a tool that can be applied in so many contexts.
It is intended to illustrate the following properties of the effective altruism approach
Being led by the heart, and not letting our heart be limited in where it takes us
George Floyd’s death might have motivated us, but that doesn’t mean that the justice system is the only place where racism is prevalent
Reason and evidence, as an expression of caring about making real change
Evidence matters because we are not devoting our time or money to these issues for tokenistic reasons, we do it because we want to effect real change
Cost-effectiveness, because nobody is a statistic
If, for the same money/time, we can help more people suffer less racism, this is great. Because every instance of racism averted counts.
The systematic approach I’ve highlighted here might seem almost callous. But to my mind it’s a sign of really caring.
Another feature of effective altruism is humility. So in that vein let me reiterate that there may be many things that I’ve got wrong here, both in terms of my thinking of what is the best thing to do for tackling racism, and in terms of capturing what EA really is.
If I were to pick one thing that’s most likely to be a mistake, it’s the omission of further research from the list of (1) direct social interventions (2) campaigning to change social norms (3) lobbying for policy change. I did this because EA isn’t just about paralysis by analysis. It’s about doing as much as it’s about thinking.
But the approach set out above is based heavily on the evidence, and if systemic racism or unconscious bias extends as far as the institutions performing the research, this jeopardises our evidence base.
And lastly our doubts about the malign influence of institutional prejudice or unconscious bias do not just reach as far as our data.
They should reach ourselves as well.
In the UK there’s a campaign group called Charity So White which tackles institutional racism within the charity sector. To finally raise a question too broad for me to answer here: Do we also need an “EA So White” too?